Skip to content

check:dev-prereqs 只判 dist 存在性,#5726 的「陈旧 dist」那半边仍无门禁 —— 而绿灯现在会提供反向保证 #5864

New issue
New issue
Closed
#6892
Closed
check:dev-prereqs 只判 dist 存在性,#5726 的「陈旧 dist」那半边仍无门禁 —— 而绿灯现在会提供反向保证#5864
#6892
Assignees
os-project-manager
Labels
domain:devxDomain lane — seat card indexed by label:pm:seatfindingRecorded observation — held, not dispatchable until the findings triage round grades it

Description

@os-zhuang
os-zhuang
opened on Aug 6, 2026

观察类发现(finding),记录自 #5795 / PR #5863 的实现过程,不建议现在就动手 —— 先留档,免得下一个人踩。

背景

PR #5863 给根 dev / dev:* 加了构建完整性前置 check:dev-prereqs:凡工作区包声明在 dist/ 下的入口不在盘上,就判「工作区未构建」并给一句 pnpm build。判据是存在性,这是刻意选的(见下)。

但 #5726 的成因是两半:

  1. packages/drivers/* 的 dist 缺失 —— PR feat(devx): 根 dev 入口新增构建完整性前置 check:dev-prereqs —— 一个前置条件一句修法 (#5795) #5863 已拦住。
  2. packages/spec/dist 陈旧(文件在,内容旧)—— 仍无门禁。这半边正是 objectstack dev 在工作区未构建时刷 12 段无关命令的 MODULE_NOT_FOUND，唯一可执行的那条却指向错误修法 #5726 里那 20+ 条「看起来非常像真实类型契约漂移」的假错误的来源:isAppResolvedDefaultToken 在 src/ 里好好导出着,只是不在陈旧的 dist 里。

为什么 PR #5863 刻意不做陈旧判定

比对 src/ 与 dist/ 的 mtime 会在每次 git worktree add 之后对所有人误报 —— 检出会重写源文件 mtime,于是 src 一律「比 dist 新」。一个开局就误报的门禁会被立刻学会忽略,比没有门禁更糟。所以 PR #5863 把陈旧明确写成非目标,并在文件头注明这半边由 AGENTS.md §9 的常备处方(pnpm install --frozen-lockfile && pnpm build)承担。

为什么仍值得记一笔

PR #5863 之前,dev 对构建状态什么都不说;之后它会打出 ✓ Workspace is built — 67 package build artifacts present.。对陈旧 dist 的场景,这行绿灯是反向保证:开发者刚被告知工作区没问题,于是更倾向于把随后的假漂移当成真 bug 去「修」—— 也就是 #5726 明确警告过的那个失败模式(在正确的代码上改出真的 bug),只是现在多了一句绿灯替它作证。

换句话说:PR #5863 让缺失那半边不可能再误导人,同时轻微加剧了陈旧那半边的误导性。净效果仍是正的(缺失是远更常见的形态),但这个副作用是新增的,不记下来就没人知道。

可能的方向(实现者自选,均未验证)

  • 构建时打戳,而不是比 mtime。 turbo 本来就为每个包算输入 hash;若 build 把它写进 dist/.build-input-hash,陈旧判定就退化成一次字符串比较,与 mtime 无关,也不受 git worktree add 影响。check-console-sha.mjs 已经是这个形状的先例(dist/.objectui-sha 对 .objectui-sha),可以照抄它的判定与「无戳则只警告」的降级。
  • 只对少数放大器包做陈旧判定(packages/spec 首当其冲 —— AGENTS.md §9 的表格里它是唯一被点名两次的),把成本压到一两个 stat。
  • 什么都不做,依赖 AGENTS.md §9 的纪律 —— 但那就该在 check:dev-prereqs 的绿灯措辞里说清它保证的是「产物存在」而非「产物最新」,别让绿灯替陈旧背书。(三条里这条最便宜,也可能就够。)

影响面

纯本地与 worktree 首启 DX,无用户可见行为变化,CI 不受影响(CI 一律先构建,且检出即新鲜,跑不出这个形态)—— 与 #5726 / #5217 同族。

参考

Activity

  1. os-zhuang
    added
    dx
    domain:devxDomain lane — seat card indexed by label:pm:seat
    findingRecorded observation — held, not dispatchable until the findings triage round grades it
    on Aug 6, 2026

  2. os-zhuang commented on Aug 6, 2026

    @os-zhuang
    os-zhuang
    on Aug 6, 2026 – with Claude
    ContributorAuthor

    发现分诊：持有（finding 保留，domain:devx 不变）

    分类：观察类。⛔ 不入 pm:queue。

    为什么持有而不是晋级

    本单与同批的 #5862 是同族但不同级，差别正在「修法是否已经确定」这一轴上：

    更硬的一条：本单自陈净效果仍是正的，坏的只是 PR #5863 顺带加剧了陈旧那半边的误导性。这是一个已知副作用的留档，不是一个待修的不变量破坏 —— 它没有「今天有人拿到错误结果」的读数，也没有 CI 判绿正确性风险，因此不适用 restore-invariant 例外。

    持有理由（一行，按 #4949 纪律）

    三条方向中两条是需要实测误报率才能定夺的设计选择，第三条是措辞级收口；建议由 devx 座位在下一次触碰 check-dev-prereqs 的单里顺带把绿灯措辞改成「产物存在」而非「产物最新」，不为此单独占一次派发。

    重看条件（不是重启条件 —— 持有态无需三要素，但记下来免得下一轮重判）

    1. check:dev-prereqs 的绿灯真的诱发了一次「在正确的代码上改出真 bug」（即 objectstack dev 在工作区未构建时刷 12 段无关命令的 MODULE_NOT_FOUND，唯一可执行的那条却指向错误修法 #5726 那个失败模式复发且有现场）⇒ 立即晋级，按 restore-invariant 处理；
    2. 或 devx 座位决定做构建时打戳（dist/.build-input-hash，形状抄 check-console-sha.mjs）⇒ 本单并入那条基建单。

    ⚠️ 一并记：本单是今日 finding 里少数**立单者主动声明「不建议现在动手」**的，本座位复核后同意 —— 这类自陈应当被信任但仍须复核，因为定级责任在分诊轮而不在立单时（#4949）。

    本评论来自分诊座位 Routine（#5474 试点），不构成认领。

    Generated by Claude Code

  3. os-zhuang
    added a commit that references this issue on Aug 6, 2026

    fix(devx): 绿灯不再声称「工作区没问题」,并把陈旧那半边的单号写进文件头

    816b7bc

  4. os-zhuang commented on Aug 6, 2026

    @os-zhuang
    os-zhuang
    on Aug 6, 2026 – with Claude
    ContributorAuthor

    分诊前先更正一处:上面三条方向里最便宜的第三条已经在 PR #5863 里落地了,别按「三条全待办」定级。

    已做:绿灯措辞收紧,不再替陈旧背书。

    - ✓ Workspace is built — 67 package build artifacts present.
+ ✓ 67 package build artifacts present (existence, not freshness).

    同时 scripts/check-dev-prereqs.mjs 的「非目标」一节写明了这个反向保证的机理并指回本单,所以缺口不再是隐含知识。

    仍待办的是前两条(真正的陈旧判定):构建时打戳(照抄 check-console-sha.mjs 的 dist/.objectui-sha 形状,与 mtime 无关)或只对 packages/spec 这类放大器做判定。本单的价值现在集中在那半边 —— 判定本身仍然不存在,pnpm dev 对陈旧 dist 依旧一言不发。

    Generated by Claude Code

  5. claude commented on Aug 7, 2026

    @claude
    claudeBot
    on Aug 7, 2026 – with Claude
    Contributor

    Findings triage (#4949 discipline): hold — finding retained, domain:devx + dx unchanged.

    Stale-premise check (origin/main @ ee3bde1): both halves of the record still read true.

    • The cheapest of the three directions is confirmed landed, exactly as the 09:56Z correction said: scripts/check-dev-prereqs.mjs:222 now prints ✓ ${checked} package build artifacts present (existence, not freshness). — the green line no longer vouches for staleness. No commit has touched that script since.
    • The remaining two directions are still unbuilt: there is no dist/.build-input-hash anywhere in the tree, and no per-package staleness assertion. pnpm dev still says nothing about a stale packages/spec/dist, which is the half that produced objectstack dev 在工作区未构建时刷 12 段无关命令的 MODULE_NOT_FOUND，唯一可执行的那条却指向错误修法 #5726's 20+ convincing fake type-drift errors.

    Why it stays held rather than promoted. The scope that remains is exactly the part the previous grading identified as design choices needing measured false-positive rates, not implementation:

    • build-time stamping (dist/.build-input-hash, shape copied from scripts/check-console-sha.mjs, mtime-independent and therefore immune to the git worktree add re-write that made mtime comparison a non-starter), versus
    • staleness assertion narrowed to a few amplifier packages, packages/spec first.

    The direction settles a build-infrastructure question — where the freshness fact is authored, and who pays for it on every dev — and the cheap wording fix that could have been a rider has already been taken. Also unchanged: this is a documented side effect, not a broken invariant. There is still no reading of "someone got a wrong result today", and no CI-green-correctness exposure, so the restore-invariant exception does not apply.

    Re-check conditions (all three unfired; restated verbatim in substance so the next round can test them without re-deriving):

    1. The green line demonstrably induces a recurrence of the objectstack dev 在工作区未构建时刷 12 段无关命令的 MODULE_NOT_FOUND，唯一可执行的那条却指向错误修法 #5726 failure mode — someone "fixes" correct code — with a concrete site ⇒ promote immediately under restore-invariant.
    2. The devx seat decides to build the build-time stamp ⇒ this folds into that infrastructure issue rather than being dispatched alone.
    3. Ownership of the freshness half moves anywhere other than AGENTS.md §9's standing prescription ⇒ re-grade.

    Related-work check, refreshed: #5862 (the same-family i18n build-prerequisite issue, promoted while this one was held) and #5726 itself are both closed/completed — so neither is a live serialization constraint on scripts/check-dev-prereqs.mjs, and the hold here is on substance, not on traffic.

    本评论来自分诊座位 Routine(#5474 试点),不构成认领。

    Generated by Claude Code

  6. os-project-manager commented on Aug 8, 2026

    @os-project-manager
    os-project-manager
    on Aug 8, 2026 – with Claude
    Collaborator

    Findings sweep (maintainer-authorized one-off, 2026-08-07 — registered on #6015): hold confirmed. The staleness half was deliberately excluded because mtime comparison false-positives on every fresh worktree — a gate that cries wolf on day one is worse than none, and AGENTS.md §9's prescription covers the gap meanwhile. Restart: a false-positive-free staleness design (e.g. content hash of spec dist) or the trap biting devs again despite §9 (note #6371 in this pool is the same trap's skill-side mitigation).

    Generated by Claude Code

  7. os-project-manager
    self-assigned this
    on Aug 9, 2026

  8. os-project-manager commented on Aug 9, 2026

    @os-project-manager
    os-project-manager
    on Aug 9, 2026 – with Claude
    Collaborator

    认领：session session_01F8q5J1MQyocgtNspb15fSn，分支 claude/issue-5864-dev-prereqs-stale-dist。

    按 #6371 的现场重新定级后开工：本单「陈旧那半边仍无门禁」的读数仍然成立，目标是给 check:dev-prereqs 补上一个误报可控的陈旧判定（方向一：构建时打戳，形状照抄 scripts/check-console-sha.mjs），并在自测里钉住「找不到输入必须红」（#4690）与一次真实陈旧树上的反向验证。

    Generated by Claude Code

  9. os-project-manager
    closed this as completedin #6892on Aug 9, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

domain:devxDomain lane — seat card indexed by label:pm:seatfindingRecorded observation — held, not dispatchable until the findings triage round grades it

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions